Skip to content

CHANGELOG.md

Latest commit

 

History

History
170 lines (113 loc) · 5.23 KB

CHANGELOG.md

File metadata and controls

170 lines (113 loc) · 5.23 KB

Version 2.15.3

  • Vectorized complex load/save.
  • Optimized complex multiplication (permissive mode).

Version 2.15.2

  • Direct translation from Symbolica.

Version 2.15.1

  • Fused load/math operations (both constant and memory access) are implemented.
  • Debugging complex codegen.

Version 2.15.0

  • Code refactoring to better support multi-threading through Applet.
  • Pool is removed as opt_level = 2 is the default.

Version 2.14.*

  • Improved inter-operability with Symbolica, especially adding external functions.
  • Addition of Applet.

Version 2.13.5

  • symjit-crate Symbolica dependency is removed.
  • Complex is re-exported.

Version 2.14.4

  • Parser complex constants parsing is fixed.

Version 2.13.3

  • get_instructions parser added.
  • Complex logical/boolean instructions changed to real only.

Version 2.13.2

  • Bridge encoder bugs fixed.

Version 2.13.1

  • order=fortran added to compile_evaluator.
  • Matrix uses fat pointers now.

Version 2.12.3

  • Mixed complex-real computation to align with Symbolica 1.3.
  • Symbolica Python interface improvements.

Version 2.12.0

  • Compiler modified to access the output of Symbolica export_instructions.
  • Translator object implements a two-pass translation of export_instructions outputs.
  • Adding compile_evaluator to the Python interface.
  • Symbolica interface supports SIMD (Application.evaluate* functions).
  • opt_level default increased to 2.

Version 2.11.0

  • Adding support for complex numbers.
  • Adding dtype to compile functions.
  • Removing explicit independent variable.
  • Fixing and consolidation stack frames between regular and fast functions.

Version 2.10.0

  • Adding support for NEON instructions on aarch64 (128-bit f64x2).
  • Adding hidden observables (named observables prefixed with __).
  • compile_json accepts regular functions in addition to ODE systems.
  • sanitize switch is deprecated.

Version 2.9.3

  • Adding support for aarch64 code segments > 1 MB (aarch64 adrp instruction).

Version 2.9.2

  • Can support >2^16 variables on aarch64.
  • Code-factoring, especially separating Expr.
  • Symbolica support added to both Python and Rust.
  • Memory leak for normal compiled code fixed (fast functions still intentionally leak memory).

Version 2.8.0

  • Explicit looping constructs Sum and Product implemented.
  • defuns keyword added, allowing for calling Python and other Symjit functions.
  • Better testing (runtests.py).

Version 2.7.1

  • An IfElse level-2 bug fixed.

Version 2.7.0

  • RISC-V machine code is supported (64-bit only).
  • IfElse explicitely supported in the low-level code genetators.
  • Improvement in register usage in the Arm64 backend.
  • A bug in Windows fast-func fixed.

Version 2.6.0

  • flush() removed and replaced with used_registers in the preamble code.
  • Alignment fixed before writing constants.
  • Level-2 optimization with improved graph coloring algorithm added.
  • opt_level argument added to compile functions.

Version 2.5.1

  • Sympy constants accepted (is_number).
  • RISC-V compilation (bytecode only).

Version 2.5.0

  • New Intemediate-Representation (MIR).
  • Peephole optimization on MIR.
  • Optional fused-mul-add instructions (fastmath option).
  • Adding apply function to Python Func.
  • Adding callable functions to Python Func.

Version 2.4.2

  • Large stack frame support in x64 (chkstk functionality).
  • Debug mode fixed.

Version 2.4.1

  • Arm64 (aarch64) large stack frame.
  • Consolidated tests suite (plots.py).
  • New intrinsic operators (Min, Max, Heaviside).
  • Switch to spec_math::cephes64 and addition of new numerical functions (erf, gamma, ...).

Version 2.4 Release Note

  • Common-subexpression elimination (keywork cse) implementede.
  • Reg and Block classes added.
  • Examples and tests updated.
  • F32 support added to low-level codegen (not exposed to the API yet).

Version 2.3 Release Note

  • Multi-threading support added (keywork use_threads).
  • Runnable class extensively rewritten to support parallelization.

Version 2.2 Release Note

  • Matrix class added to support parallelization.

Version 2.1 Release Note

  • Precise transcendental functions (expm1, log1p, exp2, and log2) added.
  • Sub-expression rewriting rules.
  • Debug mode added.
  • A subtle bug in arm code generator fixed.

Version 2.0 Release Note

  • The rust backend is extensively updated. Specially, a new register allocator improves register use.
  • The new backend uses a tree-based Intemediate Representation instead of the three-address form in version 1.
  • The default instruction set on AMD64 systems is AVX instead of SSE2 (SSE2 can still be generated if needed).
  • Improved and unified stack frame structure. The stack frame size is significantly reduced compared to version 1.
  • Fast functions (exposed rust function pointers) are introduced to reduce function call overhead of small functions.
  • Rounding functions floor and ceil are added.
  • Special fast code generation for exponentiation to an integer power (e.g., x**100).
  • Special fast code generation for modular exponentiation (e.g., x**100 % 65537).
  • Less overhead in calling standard transcendental functions.
  • Fixing a bug in comparison operators (swaping the meaning of strict vs non-strict comparisons).
  • Addition of extensive testing examples.